On Tue, 2011-05-03 at 18:08 +0200, Knut J Bjuland wrote:
is ther any plans to produce an version of freesurfer for openmpi and/or cuda 4.0? Do Anders Dale produce an version of freesurfer for openmi interface that enable use of freesurfer on supercomputer?
I would not expect a migration to CUDA 4.0 until after it has gone final (it's still on Release Candidate).
As for CPU parallelisation.... unless people want to start dealing with bigger volumes, going for MPI would probably be an overkill. OpenMP parallelisation for multicore would likely be more productive (e.g. the stallo system at notur has 8 cores per node). However, I wouldn't want to promise substantial gains there, since the memory access patterns in FreeSurfer are not uniformly optimal.
Do you need one brain processed faster, or do you just have a lot of brains to process? If it's the latter, the best speed up is to submit the jobs to your local cluster, and go on holiday. Supercomputer centres might schedule motherboards, so you'll have to have each script spin off eight copies of FreeSurfer, but the basic principle is the same.
Richard